@stableAPY

I also tried Unsloth's 35B A3B MTP on my 3060 12gb, and it's not as good as on the 3090 MTP probably gives better decode speed but leads to more experts being offloaded to the CPU so overall decode speed is lower at the end with MTP version 33 tok/s vs 39 tok/s l'll stick with classic ik_llama.cpp for my 3060

Unsloth released MTP (Multi-Token Prediction) version of Qwen 3.6 27B and 35B A3B this gives a pretty nice boost on the decode side, but it impacts a bit the prefill I think this will still be my default setup to gain a bit of decode speed, the drawback on the prefill is acceptable for me you'll need this specific branch for llama.cpp : https://t.co/cJ9sUahlqx https://t.co/k4AUdOOVKt https://t.co/6x8rL1saVv

Some MTP metadata where missing yesterday, but seems like they published the proper GGUF since

works great with thinking too https://t.co/CuHJoaVvVJ

small GPU owners, 3060's fellows, 8 or 12gb VRAM chads, > if you're using llama.cpp you should check out ik_llama.cpp it gives about 10% better performances at 128k total context running Qwen3.6-35B-A3B IQ4_XS on a 3060 12gb 31,7 tok/s at 50% of context is not that bad for a subagent for local hermes setup https://t.co/ZCN728Taoh